Re: [PATCH v2 7/8] gpu: nova-core: add NVKV typed decoding

["Eliot Courtney" <ecourtney@nvidia.com>]

On Thu Sep 17, 2026 at 5:52 PM JST, Gary Guo wrote:
> On Thu Sep 17, 2026 at 9:37 AM BST, Alexandre Courbot wrote:
>> On Thu Sep 17, 2026 at 8:53 AM BST, Gary Guo wrote:
>>> On Mon Sep 14, 2026 at 7:55 AM BST, Eliot Courtney wrote:
>>>> On Mon Sep 14, 2026 at 12:46 PM JST, Alexandre Courbot wrote:
>>>>> On Thu Aug 27, 2026 at 11:12 PM JST, Eliot Courtney wrote:
>>>>>>  use crate::gsp::nvkv::{
>>>>>> +    Array,
>>>>>>      Index,
>>>>>> +    Key,
>>>>>>      KeyId,
>>>>>>      Op,
>>>>>>      Opcode, //
>>>>>>  };
>>>>>>  use crate::num;
>>>>>>  
>>>>>> +/// Defines a schema struct together with its [`Schema`] implementation that decodes into `$target`.
>>>>>> +///
>>>>>> +/// Each member of the struct should implement `Schema`. For every (key, index, value) triple
>>>>>> +/// decoded from the NVKV stream, the generated parent `Schema` implementation will call each member
>>>>>> +/// in declaration order with that triple. If a member consumes that triple, it will stop there.
>>>>>> +/// Otherwise it will keep going until all members are tried.
>>>>>> +///
>>>>>> +/// The schema struct holds the state required by the schema implementation to do the decode. It's
>>>>>> +/// recommended to use one of the existing Schema kinds (`Required`, `Accumulated`, `Key`, `Array`,
>>>>>> +/// `Indexed`) for each member.
>>>>>> +///
>>>>>> +/// # Examples
>>>>>> +///
>>>>>> +/// ```
>>>>>> +/// nvkv_decode! {
>>>>>> +///     struct RequestSchema => Request {
>>>>>> +///         id: Required<u32, 0x0001>,
>>>>>> +///         name: Array<u8, 64, 0x0002>,
>>>>>> +///     }
>>>>>> +/// }
>>>>>> +/// ```
>>>>>> +macro_rules! nvkv_decode {
>>>>>> +    (
>>>>>> +        $(#[$attr:meta])*
>>>>>> +        $vis:vis struct $name:ident => $target:ident {
>>>>>> +            $(
>>>>>> +                $(#[$field_attr:meta])*
>>>>>> +                $field_vis:vis $field:ident : $ty:ty
>>>>>> +            ),* $(,)?
>>>>>> +        }
>>>>>> +    ) => {
>>>>>> +        $(#[$attr])*
>>>>>> +        $vis struct $name {
>>>>>> +            $(
>>>>>> +                $(#[$field_attr])*
>>>>>> +                $field_vis $field: $ty,
>>>>>> +            )*
>>>>>> +        }
>>>>>> +
>>>>>> +        impl $crate::gsp::nvkv::Schema for $name {
>>>>>> +            type Target = $target;
>>>>>> +
>>>>>> +            fn init() -> impl ::kernel::prelude::Init<Self> {
>>>>>> +                ::pin_init::init!(Self {
>>>>>> +                    $( $field <- <$ty as $crate::gsp::nvkv::Schema>::init(), )*
>>>>>> +                })
>>>>>> +            }
>>>>>> +
>>>>>> +            fn visit(
>>>>>> +                &mut self,
>>>>>> +                key: $crate::gsp::nvkv::KeyId,
>>>>>> +                index: $crate::gsp::nvkv::Index,
>>>>>> +                value: $crate::gsp::nvkv::DecoderValue<'_>,
>>>>>> +            ) -> ::kernel::error::Result<bool> {
>>>>>> +                Ok(false
>>>>>> +                    $( || $crate::gsp::nvkv::Schema::visit(&mut self.$field, key, index, value)? )*)
>>>>>
>>>>> Mmm looks like this is going to be `O(n)` with `n` being the number of
>>>>> fields?
>>>>>
>>>>> This is ok for a first implementation but eventually I hope we can
>>>>> switch to a more efficient dispatch.
>>>>
>>>> I thought quite a bit about this while writing this code, since we need
>>>> the escape hatch to imperative decode (custom Schema impl basically). To
>>>> be able to get it down to a match on the key, we need to know ahead of
>>>> time which keys a Schema will consume. That duplicates the info from the
>>>> visit() implementation.
>>>>
>>>> I thought up a few methods but it's unclear to me which one is best, so
>>>> I just left it for now. Please LMK if you think this is urgent, I can
>>>> try in a follow up to improve this. Here are my ideas (when I say O(1)
>>>> lookup I mean modulo how the compiler decides to do it with the set of
>>>> key IDs it gets):
>>>>
>>>> 1. current code - just visit()
>>>> pros: key source of truth not duplicates
>>>> cons: O(field) visit as you say
>>>>
>>>> 2. Associated const KEY_ID: Option<KeyId> - None if a Schema accepts multiple keys.
>>>> You can match on each associated const in the macro.
>>>> pros: O(1) if the current key goes to a field with KEY_ID = Some(...)
>>>> cons: O(#fields accepting multiple keys) if current key is one of them
>>>>
>>>> 3. fn accepts() -> bool
>>>> You can match on `if F::accepts(key)` for each field. We could potentially make
>>>> this const with Gary's const traits polyfill.
>>>> pros: O(1) if you write an inline-able+optimizable implementation.
>>>>
>>>> 4. Associated const KEYS table; use tricks to concat tables
>>>> pros: O(1) lookup 
>>>> cons: actually MSRV can't get this to optimize down to O(1) 
>>>>   if you use slice::contains(), but stable can.
>>>
>>> Hmm, am I missing the obvious? Why not generate a `match` expression on IDs of
>>> fields? It looks like in the example all keys would have a known ID to the
>>> macro.
>>
>> Some keys may come from embedded structs, which the macro has no way to
>> see.
>
> You can match all keys that you can see, and delegate to embedded structs if
> keys are not known.
>
> This is essentially the same pattern that `#[serde(flatten)]` uses, just
> replacing identifier names with keys.
>
> The keys don't need be part of the type system, and it just additional metadata
> for the macro to generate correct impl.
>
> Best,
> Gary

This sounds approximately equivalent to the option #2 I mentioned
(w.r.t. delegating if the keys are not known).

I considered a serde-like approach where the metadata is annotated for
the macro rather than by the type system, but I wanted to keep the macro
simple and not introduce more macro-DSL than I need to. And also keep it
kinda easy to write manual Schemas by not keeping the deserialization
info in macro metadata.